Utilizing Weakly Trained GANs to produce 3D Models


Generative Adversarial Networks (GAN) have become the de-facto standard for generative models. However large amounts of data are required to achieve excellent performance which makes GANs challenging to apply to sparse datasets. In order to utilize GANs with sparse datasets, we propose using weakly trained GANs to produce models which we then use to augment existing data. For this study, we utilize 2 classes of a popular 3D dataset called 3DShapeNet to train on. Next data augmentation is performed on the original models using our proposed method, lastly, we conduct a study to assess our models on Amazon Mechanical Turk.

Download full paper here